OpenAI released GPT-6 Cyber on 5 November 2026 to a limited group of approved security customers, rating it Critical for cyber capability. Two days earlier, the UK's National Cyber Security Centre judged that many systems will remain vulnerable to AI cyber tools by 2028. We read both as a reason to review patching speed and ageing systems.
What happened
5 November, the release. The Register reports that GPT-6 Cyber is initially available only through the API and Codex, to selected customers in OpenAI's Daybreak Red programme, which covers authorized work such as penetration testing and red teaming. OpenAI plans to widen access to all Daybreak Red customers over the coming weeks. Security Boulevard adds that safeguards include IP allowlisting and monitoring. This is not a general business product.
The rating. OpenAI classes the model as Critical, the highest cyber tier in its Preparedness Framework. The Register explains that this tier covers models able to develop working zero-day exploits in real-world critical systems and to carry out end-to-end attacks. Security Boulevard notes that the rating is not a safety certification.
OpenAI's claimed results. These are OpenAI's own figures, and none of the coverage we read cites an independent evaluation:
| Test | OpenAI's claim |
|---|---|
| Firefox | 81 previously unknown flaws, which OpenAI says Mozilla classed as high severity and has started fixing |
| Linux kernel | 948 vulnerabilities with working exploits, 588 of them previously unknown |
| ExploitGym benchmark | 48.5%, against 32.1% for GPT-5.6 Cyber |
| Staying in scope in tests | 99.93% without deployment safety prompts, and every tested case with them |
The Register reports that OpenAI declined to give further details on the 0.07% of cases that went beyond the intended scope, including whether any involved real-world incidents.
3 November, the NCSC assessment. The NCSC published "Artificial intelligence and cyber intrusion", which looks at threats to 2028. It judges it highly likely that AI-powered malware will have been used in a ransomware attack on UK critical national infrastructure (CNI) by then, and likely that attackers will apply AI cyber capability against known vulnerabilities in CNI legacy technology. It also judges it almost certain that many systems will remain vulnerable to AI cyber tools. The report's scenarios are described as plausible possibilities, not predictions.
What it means for businesses
The NCSC's ransomware and legacy-technology judgements concern UK critical infrastructure, although its judgement that many systems will stay vulnerable is not limited to it. GPT-6 Cyber is restricted to approved customers, and none of the coverage we read reports it being used in an attack. Three practical points follow, and these are our reading, not statements by either organization.
Expect more patches, and sooner. If AI tools find flaws at anything like the claimed scale, vendors will have more fixes to ship and customers more updates to apply. In its baseline scenario for 2028, the NCSC writes that "an organisation's internal patching and testing processes will need to evolve to meet the pace of accelerating vulnerability discovery." A patch process that takes a month may not keep up.
Old systems have fewer ways out. A flaw in supported software can usually be fixed with an update. A system that no longer receives updates, or that nobody dares to change, has no such route. The NCSC's concern about known flaws in legacy technology is specific to CNI, but we think the same logic applies to an old ERP or a line-of-business application.
Defence is not automatic. In its baseline scenario for 2028, which the NCSC says is not a prediction, defensive AI tools shorten the time to a patch, but many systems stay vulnerable without correct and widespread configuration. Buying a tool does not close the gap on its own.
What to do next
- List your software and its support status. Include custom applications, third-party libraries and anything internet-facing. Name an owner for each item and record when it last received a security update.
- Time your patch process. Measure the days from a vendor fix to live deployment, including testing before release, then shorten the slowest step.
- Decide what to do with unsupported systems. Isolate them from the internet, replace them, or modernize them in stages. Our legacy modernization assessment checklist helps set the order.
- Put patching into contracts. A software service level agreement can set response times for security fixes and say who applies them.
- Keep Firefox and Linux servers on current updates. Mozilla has started fixing the reported Firefox flaws, so we see those updates as the most direct follow-on. The Linux findings are still with maintainers as proposed patches, so watch for kernel updates from your distribution.
- Set rules for AI-assisted security testing. If a supplier offers it, agree in writing which systems are in scope, who authorized the work, and who reviews the findings. OpenAI's own tests were not perfect without safety prompts, and it declined to detail the exceptions. We read that as a reason to keep scope under review, not as evidence about any supplier's tool.
- Treat vendor benchmarks as claims. Ask what was tested and whether anyone outside the vendor has confirmed the results.
Timeline Digital, a custom software development company founded in 2013 in Islamabad, builds and supports business software as managed projects, including legacy system modernization and software maintenance and support. Every project starts with a free pilot: Before you commit to the full project, we build 2 to 3 of your key modules as working software, free of charge, ready in about 2 days. Your team tests the pilot, and the full build starts only after you approve it.
Sources
- OpenAI releases GPT-6 Cyber, and promises this one stays within 'rules of engagement', The Register, 5 November 2026
- OpenAI Launches GPT-6 Cyber to Find Security Flaws, Limits Access Over Hacking Risks, Security Boulevard, 6 November 2026
- OpenAI Launches Critical-rated GPT-6 Cyber For Approved Security Teams, Dataconomy, 6 November 2026
- Artificial intelligence and cyber intrusion, National Cyber Security Centre, 3 November 2026